A new research paper evaluates the robustness of foundation models when applied to mammography, particularly under domain shift conditions. The study used a standardized protocol to test 15 different foundation model backbones across various datasets, assessing their performance on breast density, BI-RADS severity, and cancer status. While mammography-specific models like Mammo-FM and MaMA showed strong out-of-distribution performance, their robustness was not solely dependent on mammography exposure. The research highlights the importance of dataset-level evaluation for assessing mammography representations and notes that even leading models exhibit varied performance across different datasets. AI
IMPACT Highlights the need for robust foundation models in medical imaging and identifies key areas for future development in AI for mammography.
RANK_REASON Research paper published on arXiv detailing model benchmarking. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
